Foundation Model

A foundation model is a single model, pretrained once on broad data at large scale, designed to be adapted into many different downstream applications rather than built for one specific task from the start. Bommasani and colleagues at Stanford's Center for Research on Foundation Models coined the term explicitly in their 2021 paper "On the Opportunities and Risks of Foundation Models," naming the pattern they saw emerging around models like GPT-3 — released by Brown and colleagues the year before, with 175 billion parameters trained on roughly 300 billion tokens — which could be adapted through prompting or light fine-tuning into translation, summarization, coding, and dozens of other applications nobody specifically trained it to do, all from the same underlying weights. The word "foundation" is doing real work in the name: everything built on top inherits whatever the base model got right or wrong.

The "foundation" framing is a structural warning as much as a description, and that's the paper's real point: a flaw in the base propagates into everything built on it. Because dozens or hundreds of separate applications can all sit on top of the identical pretrained weights, a bias, a factual gap, or a security vulnerability present in the foundation model doesn't stay contained to one deployment — it's inherited by every single downstream system built from that base, including ones whose developers never touched the original training data and have no way to fix the underlying issue themselves. This is structurally different from the old approach of training a separate model per task, where a flaw in one model's training data only ever affected that one model.

Scale alone doesn't make a model a foundation model — the defining property is breadth of downstream reuse, which is why the term applies across modalities, not just to large language models. CLIP (OpenAI, 2021) is a foundation model for vision-language tasks, adapted into zero-shot image classification and image search without ever being retrained for either; Whisper (OpenAI, 2022) is a foundation model for speech, reused across transcription, translation, and voice-interface products built by teams who never touched its training data. A model trained once and reused across genuinely different applications and teams qualifies under Bommasani and colleagues' definition, whether that model happens to be a 175-billion-parameter language model or a much smaller speech or vision network, as long as the "pretrain once, adapt broadly" pattern holds.

Governance and accountability for a foundation model are harder than for a single-purpose model precisely because responsibility is distributed across a chain nobody individually controls end to end. The team that trains the base model doesn't know every downstream use it will be adapted to; the team building a specific downstream application often can't inspect or retrain the base weights themselves. This split is exactly why foundation-model providers increasingly publish usage policies, model cards, and evaluation results at the base-model level — it's an attempt to give every downstream adopter the information needed to judge whether the foundation they're building on actually fits their specific application, since no downstream team can fully audit a base model they didn't train.

The word "foundation" names a trade-off as much as a technique: enormous efficiency from training once and reusing broadly, paired with a genuinely new kind of systemic risk where the base model's limitations don't stay isolated to one product but spread silently through every single application resting on it — which is exactly the opportunity-and-risk pairing Bommasani and colleagues' 2021 paper set out to name in the first place.